An Adaptive Edge-Guided Dual-Network Framework for Fast QR Code Motion Deblurring

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional image deblurring prioritizes visual quality, not decoding accuracy.
• Existing QR code deblurring methods often implicitly learn structural information.
• QR codes require preservation of critical structures for successful decoding.

Purpose of the Study:

• To develop a novel approach for effective QR code deblurring that prioritizes decoding success.
• To explicitly incorporate structural priors into the deblurring process.
• To create an adaptive system balancing restoration accuracy and computational efficiency.

Main Methods:

• Proposed an Edge-Guided Attention Block (EGAB) to explicitly extract and utilize edge information.
• Developed an Edge-Guided Restormer (EG-Restormer) for severe blur and a Lightweight and Efficient Network (LENet) for mild blur.
• Integrated EG-Restormer and LENet into an Adaptive Dual-network (ADNet) for blur-level-based restoration selection.

Main Results:

• EG-Restormer improved QR code decoding rates by 8.67 percentage points.
• The proposed methods achieved the highest decoding rates after fine-tuning on QRData.
• ADNet reduced average inference latency by 19% with comparable decoding performance.

Conclusions:

• Explicitly modeling edge priors enhances the recovery of decoding-critical structures in QR codes.
• Adaptive network routing effectively balances QR code deblurring accuracy and computational efficiency.
• The proposed methods offer a significant advancement in robust QR code restoration.